eval-iterate

# Eval Iterate Drive a failing AI/LLM eval to a real green: diagnose, fix, re-run, confirm — capped at 5 iterations, never by weakening the eval. This `SKILL.md` is the **orchestration index**. Load the matching rule file when you need detail — do not preload them. | Phase | Goal | Required rule | | ----- | ---- | ------------- | | 0 | Resolve the target eval + capture the baseline failure | this file | | 1 | Resolve how to run it | this file | | 2 | Classify the failure (verdict required) | [`rules/eval-bug-classification.md`](./rules/eval-bug-classification.md) | | 3 | Apply the minimal fix — gated if it touches the eval itself | [`rules/anti-gaming-guard.md`](./rules/anti-gaming-guard.md) | | 4 | Re-run, then confirm with a second run | [`rules/convergence-confirmation.md`](./rules/convergence-confirmation.md) | | 5 | Iterate or stop at the cap | this file | | 6 | Report (structured exit summary) | this file | Always read [`rules/anti-gaming-guard.md`](./rules/anti-gaming-guard.md) before touching any eval definition (assertion, threshold, golden-set item, judge prompt). The refusals in it apply on every iteration. ## Input The user provides one of: - An eval identifier — an L2 suite name (e.g. `tier-routing`), a golden-set file path, or a test/eval file path. - A PR URL with a failing eval check. - **Nothing** — if `$ARGUMENTS` is empty, auto-detect the failing eval check on the current branch's open PR (see Phase 0). - `--max-iterations <n>` — lowers the cap b...